ToolNest AI

Unsloth AI

Open-source fine-tuning & reinforcement learning for LLMs. 🦥

Visit Website
Unsloth AI

What is Unsloth AI?

Unsloth makes it super easy for you to train text-to-speech (TTS), diffusion, multimodal/image and text models like Llama 3 100% locally or for free on platforms such as Google Colab and Kaggle.

We streamline the entire training workflow, including model loading, quantizing, training, evaluating, running, saving, exporting, and integrations with inference engines like Ollama, llama.cpp, and vLLM.

How to use

Unsloth can be installed locally via Linux, Windows, Kaggle, or another GPU service like Google Colab. Most use Unsloth through the interface Google Colab which provides a free GPU to train with.

Core Features

  • Text-to-speech fine-tuning (advanced voice cloning)
  • Vision fine-tuning
  • Accurate Dynamic quantized models for you to run
  • Bug fixes for open models

Use Cases

  • More accurate and advanced voice cloning
  • Customized models for personal chatbots, characters, personalities etc.
  • RL for domain specific use-cases like law, medicine, finance

FAQ

Is Unsloth Free and open-source?
Yes Unsloth is fully open-source and free to download locally on your own device. You do not need to connect to the internet or any API while using Unsloth.
Does Unsloth have a paid offering?
At the moment Unsloth does not have a paid offering. The product is free to use.

Pricing

Pros & Cons

Pros
  • Significantly faster training (up to 30x faster than FA2)
  • Substantially reduced memory usage (up to 90% less than FA2)
  • Open-source and beginner-friendly
  • Supports a wide range of LLMs and GPU types (NVIDIA, AMD, Intel)
  • Offers faster inference capabilities
  • More energy-efficient and environmentally friendly
  • Enables rapid custom model training (e.g., 24 hours vs 30 days)
Cons
  • MultiGPU support for the free version is still 'coming soon'
  • Pricing for Pro and Enterprise plans requires direct contact with the company
  • The 'even faster inference' feature is still 'in the works', implying current inference might not be at its peak potential